
Opponent Modeling in Stratego J.A. Stankiewicz 24th June 2009 Abstract the player might become suspicious that the opponent's 1 piece is stronger than his own. Stratego is a game of imperfect information, However, it might be the case that the opponent is where observations of the opponent's behaviour bluffing, pretending the piece is stronger or weaker than are crucial for determining the best move. This it really is. Keeping track of the opponent's bluffing be- paper describes how one can model the oppo- haviour in past situations gives valuable information for nent in the game of Stratego, using a Bayesian the future. If an opponent is known to bluff in a certain approach. By observing the moves of the op- situation, one can take advantage of this by making a ponent, a probability distribution can be de- different move than one normally would. rived to help determine the identity of unknown The focus of this paper is on how to model the op- pieces of the opponent. Experiments show that ponent in Stratego. Certain techniques are translated there is a significant increase in the percentage from or inspired by Poker, a game in which some re- of correctly guessed unknown pieces. More- search regarding opponent modeling has been done al- over, the average probability assigned to the ready [3,5,14]. The goal is to create an opponent model real identity of an unknown piece, shows an in- which improves the performance of a Stratego playing crease as well. These results eventually trans- agent significantly. late into an improved win rate. The structure of this paper is the following. Section 2 gives some background on the game Stratego and its ba- 1 Introduction sics. Section 3 then gives a short summary of previous research done on opponent modeling. Section 4 explains Stratego is a turn-based two-player game of imperfect the methods used for creating a model of the opponent in information, played on a 10×10 grid. Each player starts Stratego. The experiments and results using these meth- with 40 pieces, each piece having a certain rank (see ods are discussed in Section 5. Finally Section 6 draws Section 2). The goal of the game is to either capture the the conclusions and suggests areas for future research to opponent's flag or to capture enough pieces such that improve the opponent model. the opponent cannot make any moves. The player has to deal with pieces of the opponent, some of which the ranks are unknown. In fact, at the 2 Overview of Stratego start of the game, none of the opponent's pieces are Stratego was developed by Mogendorff in 1942. The known. In this case, one can use heuristics to determine game in its present form first appeared in 1961, licensed an initial probability distribution for every piece, based by the Milton Bradley Company. The game is based on setup statistics from a database of games [7, 8]. Pre- on the game `L'Attaque' which was invented by Mdm. dicting the probability on each rank for every unknown Hermance Edan and dates back to 1908. Many different piece can greatly help in determining the best move to variants of Stratego were released throughout the years, make. In order to make an accurate prediction, it is nec- such as 3-player and 4-player games. This paper how- essary to model the opponent's playing style, based on ever, uses the original 2-player variant as its basis. the game's history. The process of creating such a model is called opponent modeling. 2.1 Ranks and Movement Each time the opponent makes a move, it may pro- The rules described in this section are an edited version vide new information on the rank of an unknown piece. of the rules published in 1986 by the Milton Bradley For instance, if the opponent moves an unknown piece Company [11]. towards a player's piece of which he knows the rank, then At start, each player has 40 pieces at his disposal to place in a 4×10 area on the board. There are 12 ranks 1Stratego is a registered trademark of Hauseman & Hotte N.V., Amsterdam, Netherlands. All Rights Reserved. c 2002 Hasbro, in total. These are listed in Table 1, together with the Pawtucket, RI 02862. amount of pieces with that rank. J.A. Stankiewicz Name Quantity The most important piece is the Flag, so its initial Marshal 1 position and surrounding pieces should be chosen wisely. General 1 Most players choose to place the Flag on the very last Colonel 2 row and to surround it with some Bombs, while putting Major 3 the other Bombs at a different location to let the oppo- Captain 4 nent think the Flag is somewhere else. Another way to Lieutenant 4 bluff about the Flag is to place it where the opponent Sergeant 4 does not expect it, like more on the front lines. This of Miner 5 course is a risky strategy. Scout 8 The placement of the other pieces is also important. Spy 1 Placing the high-ranked pieces on the front lines will re- Bomb 6 veal their location quickly and therefore it is better to Flag 1 avoid it, since these pieces become of greater value to- wards the end of the game. However, it is recommended Table 1: Stratego Ranks and Quantities to put some high-ranked pieces on the front lines, in order to have some strength there as well. The Scout, Miner and Spy are pieces that become The ranks are listed from strongest to weakest, ex- more valuable towards the end as well. Since the Scout cept the Bomb and Flag which are special ranks. Pieces can move multiple squares, it can be used to detect bluffs may move one square at a time, horizontally or vertically. and Bombs when the board is less crowded. The Miner In the middle of the board, there are two lakes of size is essential in order to get to the Flag, as the Flag will 2×2 where no piece is allowed to move to. A player may probably be surrounded by Bombs. The Spy becomes attempt to capture an opponent's piece if the player's more valuable because it is the only piece, except for piece begins its turn on a square orthogonally adjacent the Bomb and the Marshal, that is able to capture a to the opponent's piece. In general, a piece is captured Marshal. if its rank is lower than the rank of the piece it is attack- A good strategy is one where the player gives away ing or attacked by. If the rank of the two pieces is equal, the least information about his pieces, by either con- both pieces are captured. There are some exceptions to cealing their identity as long as possible, or giving his these rules however. opponent false information by making certain moves. It The Scout may move multiple squares in a straight is best to capture pieces using a piece with a rank that is line and is allowed to attack a piece at the end of his not much higher than that of the piece it captures. This path, like the rook in Chess. The Bombs and the Flag way the player does not unnecessarily reveal the position on the other hand, are the only immovable pieces in the of high-ranked pieces. game. The Spy can capture the Marshal, if the Spy is the attacking piece, making it the only movable rank (except 3 Related Research for the Marshal itself) that is able to capture the Mar- Opponent modeling in Stratego is a fairly new topic. shal. If the Marshal attacks the Spy however, the Spy is Most research regarding opponent modeling has been captured. focused on Poker, which is a non-deterministic game of The Bomb captures any rank, except for the Miner, imperfect information. Different approaches have been which is the only rank able to capture Bombs. If a Bomb used to model the opponent. In 1998, Billings et al. [3,4] is attacked by a rank other than the Miner, the Bomb introduced a poker playing agent, which used weights stays on the board. to determine the likeliness of its opponents holding a The game ends when a player captures the Flag of certain pair of cards. These weights were modified ac- his opponent or one of the players cannot make any more cording to the observations of the community cards and moves. If none of the players can make any more moves, the opponent's actions. In 2000, Davidson et al. [5, 6] the game ends in a draw. proposed an opponent modeling method using artificial neural networks. The network used a set of inputs to de- 2.2 Strategy Basics termine the opponent's next action given these inputs. As the name of the game suggests, a good strategy is Korb et al. [10] proposed a poker playing agent which the key to success. Such a strategy has to be deter- uses a Bayesian network to model its opponent. Other mined right from the start of the game, when setting Bayesian opponent modeling methods were proposed by up the pieces in their initial positions. Below is a short Southey et al. [14] in 2005 and Ponsen et al. [12] in 2008. summary of strategies described by De Boer [7]. In 2007, De Boer [7,8] introduced a Stratego playing (v. June 18, 2009, p.2) J.A. Stankiewicz agent. This agent uses a basic opponent model.
Details
-
File Typepdf
-
Upload Time-
-
Content LanguagesEnglish
-
Upload UserAnonymous/Not logged-in
-
File Pages8 Page
-
File Size-